Egocentric and Exocentric Data Collection

Egocentric and exocentric, captured together

Every session records both viewpoints against the same clock, so the pair can be used as a pair rather than reconciled after the fact.

• Head-mounted first-person capture, framed and stabilised for the task rather than for general wearability.
• External multi-view rigs positioned to cover the hands, the manipulated object, and the actor's body pose the head camera cannot see.
• Timecode synchronisation across every stream, so egocentric and exocentric footage line up frame to frame rather than by estimate.
• Calibration recorded with the footage — camera intrinsics and rig geometry captured at the time of the session, not reconstructed afterwards from what the frames happen to show.

Why the paired view is worth more than either alone

The egocentric stream shows what the actor saw: the approach, the viewpoint the robot's own camera will eventually replicate, the visual cues the actor was reacting to. It also shows very little of what the actor's hands were actually doing, because a head-mounted camera looks where the head points, and hands, tools, and the object being manipulated spend much of a task out of that frame or blocked by the actor's own arm. The exocentric stream shows the opposite: hand articulation, grasp, object pose, and body posture, all visible from a fixed external angle, but without the viewpoint a policy trained for first-person input needs. Neither stream is a substitute for the other. Watched alone, either is informative to a person. Paired and synchronised, the two together give a learning system what it actually needs — the viewpoint it will act from and the physical detail that viewpoint alone omits. EgoExo4D, a public research dataset, is the reason this pairing is now standard practice in embodied AI rather than a bespoke idea: it demonstrated that synchronised ego-exo capture produces trainable data that neither stream produces on its own. It built on Ego4D, the public dataset that established egocentric video as a research field in the first place, by adding the exocentric half most training use cases actually need.

How a session runs

The variables that ruin a capture session are rarely visible until playback, which is why the sequence is built to catch them before the crew leaves.

• Calibration before the first take, with every camera and rig checked against the session's reference markers, not assumed from the previous session.
• Activity scripting written in advance, so every take of the same task follows the same steps in the same order and takes are comparable to each other.
• Operator briefing on the task, the rig, and the failure modes specific to that session, before recording starts.
• In-session quality checks — footage reviewed between takes, not batched up for a single pass at the end of the day.
• Same-day review before the crew stands down, because a reshoot is inexpensive on day one of a programme and effectively impossible on day thirty, once the set, the props, and the actors are gone.

Teleoperation demonstrations

We support teleoperation capture when the client supplies the rig — the robot arm, the controller, the manipulation setup are yours; the operator training, session discipline, and take-level quality control are ours. Teleoperation demonstrations of this kind are the standard training input for vision-language-action (VLA) models, and that downstream use is what sets the economics here apart from scripted human activity capture: demonstration counts per manipulation task for a usable training set typically run into the hundreds, not the dozens, so throughput planning and session repeatability matter more here than polish on any individual take. We plan the programme around that volume from the first conversation rather than discovering it partway through.

What you receive

Delivery is built to enter a training pipeline directly, not to be reformatted after the fact.

• Synchronised video for every stream captured in the session — egocentric and every exocentric angle, aligned to a common timecode.
• Calibration and pose metadata for every camera and rig used, delivered alongside the footage rather than as a separate reconstruction step.
• Per-session logs tying individual takes back to the script step and task variant they belong to.
• Annotation on the same footage when you want it — the labelling work described on our data annotation page runs directly against what this page captures.

• Data annotation
• AI training data